Anthropic shares Claude results on ExploitBench, with Opus 5 leading key safety metrics
TheZvi · x · 2026-07-25
Anthropic shares Claude results on ExploitBench
The post highlights Anthropic’s latest results on ExploitBench, a benchmark for measuring how models handle exploit-style tasks and sandbox-escape attempts.
The figure shared in the post compares several Claude models on the benchmark and shows:
- Claude Opus 5 with the highest AutoNudge Mean among the listed Claude models
- Claude Opus 5 also leading in AutoNudge Cap% and Full ACEs
- Older or smaller Claude variants scoring lower across the same metrics
The benchmark description says the evaluation was run across multiple trials and environments, and that Full ACEs refers to complete exploits that achieve arbitrary code execution combined across both plain and AutoNudge settings.
More from Safety
- AI shrinks the patch window to hours; air-gapped backups are your last line — aakashgupta · 2026-10-26
- NYT roundtable asks five experts what the rules for AI companies should actually be — ShakeelHashim · 2026-10-26
- Google DeepMind pilots world's first double-blind evaluations for frontier models — scychan_brains · 2026-10-26
- LLM back-of-envelope estimate puts PHP vulnerability impact at $1bn–$4bn — S_OhEigeartaigh · 2026-10-26
- US states halt new data centers as 72% of Americans oppose local builds — YvesMulkers · 2026-10-26
- LEMON: zero-shot LLM misbehavior monitor flags hallucinations from attention maps — yoavartzi · 2026-10-26